Self-Adapting Language Models

Adam Zweiger (MIT) · Jyo Pari (MIT) · Han Guo (Massachusetts Institute of Technology) · Yoon Kim (Massachusetts Institute of Technology) · Pulkit Agrawal (MIT)
adaptation modulesauxiliary networksdata augmentationdownstream performancefew-shot generalizationfinetuning datagradient-based updatesknowledge incorporationoptimization hyperparameterspersistent weight updatesreinforcement learning loopself-adapting llmsself-editsupervised finetuningupdate directives

Large language models (LLMs) are powerful but static; they lack mechanisms to adapt their weights in response to new tasks, knowledge, or examples. We introduce $\textbf{Se}$lf-$\textbf{A}$dapting $\textbf{L}$LMs (SEAL), a framework that enables LLMs to self-adapt by generating their own finetuning data and update directives. Given a new input, the model produces a $\textit{self-edit}$